Skip to content

v1.2.1: Compaction Memory Quality — Four-Layer Defense

Choose a tag to compare

View all tags
@sdwolf4103 sdwolf4103 released this 26 Apr 09:07
· 94 commits to main since this release
v1.2.1
6603fe8

v1.2.1 (2026-04-26)

Compaction Memory Quality — Four-Layer Defense

This release hardens Workspace Memory quality after Memory V2 by fixing the main failure modes seen in real compaction output: silently lost memories, duplicate facts, stale entries, and session snapshots leaking into long-term memory.

Four-Layer Defense

  • Prompt: Durable-content guidance keeps the model focused on factual, long-lived memories
  • Parser: Accepts bracketless format and filters session-snapshot noise
  • Storage: Entity-key deduplication, topic supersession, and source-priority-aware selection
  • Staleness: Age-based pruning cleans obsolete compaction/manual entries automatically

Key Fixes

  • P0a: Accept bracketless memory format (text not found) so real LLM output no longer gets silently dropped
  • P0b: Eliminate purple/italic rendering issues by replacing fragile wrapper formats with durable plain-text/Markdown-safe output
  • P0c: Reject session-snapshot project memories such as file counts, test counts, and Phase progress
  • P0d: Add self-cleaning storage behavior with deterministic deduplication, supersession, and stale-entry pruning

Additional Improvements

  • Short but important reference facts like Admin PIN and config values are now preserved
  • Chinese/English snapshot filtering is more robust, including
  • Feedback and decision memories now prefer newer facts when they supersede older verbose entries

Verification

  • 70/70 tests pass
    This release is the hardening layer on top of Memory V2, making long-term memory more trustworthy in day-to-day use.
    Full release notes: RELEASE_NOTES.md
Assets 2
Loading